AI tools behaving badly — like Microsoft’s Bing AI losing track of which year it is — has become a subgenre of reporting on AI. But very often, it’s hard to tell the difference between a bug and poor construction of the underlying AI model that analyzes incoming data and predicts what an acceptable response will be, like Google’s Gemini image generator drawing diverse Nazis due to a filter setting.
Now, OpenAI is releasing the first draft of a proposed framework, called Model Spec, that would shape how AI tools like its own GPT-4 model respond in the future. The OpenAI approach proposes three general principles — that AI models should assist the developer and end-user with helpful responses that follow instructions, benefit humanity with consideration of potential benefits and harms, and reflect well on OpenAI with respect to social norms and laws.
It also includes several rules:
OpenAI says the idea is to also let companies and users “toggle” how “spicy” AI models could get. One example the company points to is with NSFW content, where the company says it is “exploring whether we can responsibly provide the ability to generate NSFW content in age-appropriate contexts through the API and ChatGPT.”
Joanne Jang, product manager at OpenAI, explains that the idea is to get public input to help direct how AI models should behave and says that this framework would help draw a clearer line between what is intentional and a bug. Among the default behaviors OpenAI proposes for the model are to assume the best intentions from the user or developer, ask clarifying questions, don’t overstep, take an objective point of view, discourage hate, don’t try to change anyone’s mind, and express uncertainty.
“We think we can bring building blocks for people to have more nuanced conversations about models, and ask questions like if models should follow the law, whose law?” Jang tells The Verge. “I am hoping we can decouple discussions on whether or not something is a bug or a response was a principle people don’t agree on because that would make conversations of what we should be bringing to the policy team easier.”
Model Spec will not immediately impact OpenAI’s currently released models, like GPT-4 or DALL-E 3, which continue to operate under their existing usage policies.
Jang calls model behavior a “nascent science” and says Model Spec is intended as a living document that could be updated often. For now, OpenAI will be waiting for feedback from the public and the different stakeholders (including “policymakers, trusted institutions, and domain experts”) that use its models, although Jang did not give a timeframe for the release of a second draft of Model Spec.
OpenAI did not say how much of the public’s feedback may be adopted or exactly who will determine what needs to be changed. Ultimately, the company has the final say on how its models will behave and said in a post that “We hope this will provide us with early insights as we develop a robust process for gathering and incorporating feedback to ensure we are responsibly building towards our mission.”